Build your first workflow

Start a blank workflow

Open Workflows and press New workflow: an empty canvas opens. Click its title in the bar at the top, New orchestration, and name it Answer a customer question. Open Workflow settings and, under Description, describe it in a line: Answers a stockist's question from the compliance guides and our records, with citations. The description is what people see in the list, and what Synapse AI reads when it is asked to use this workflow.

Add the search

Press Add step in the dock at the bottom of the canvas and click Retrieve, or drag it onto the canvas. The first step you add is where a run starts, marked START; Set as entry point in a step's setup card moves that mark.

Click the step to open its setup card on the right and set it up:

• Name — Search the guides.
• Knowledge bases — pick the bases to search: Compliance guides and Suppliers & invoices.
• Query — what to search for. The run's input lands in state.user_input, so type {{ state.user_input }}; typing {{ lists what you can use.
• Output key — sources. The passages land there, for the next step to read.

A Retrieve step makes no model call; it only searches, so it costs a fraction of a cent.

Add the answer

Add an LLM step the same way — typing in Add step's list finds one by name — and connect it: drag from the right-hand handle of Search the guides to the left-hand handle of the new step, or choose the new step under Next step in the first one's card. Then set it up:

• Name — Answer.

• Prompt template — the whole prompt, with state filled in where you write {{ state.… }}:

  Answer the stockist in under 120 words using only the passages below, citing [K#].
  If they do not cover it, say so and offer the order desk.
  
  Question: {{ state.user_input }}
  
  {{ state.sources.context }}

  sources.context is the passages the search found, each under its label — [K1], [K2] — ready to put in a prompt.

• Model override — leave it on the workspace default, or pick one for this step.

• Output key — answer.

Say what the run makes

Open Workflow settings and, under Guardrails, set Result to answer. The result is what the run hands back: the reply in chat, the response over the API, and what a rubric scores. Left on its default, it is the last step's output, which here is the same; naming it keeps it right when you add a step later.

Its State Schema lists the keys the workflow uses — sources and answer; give each a description for the next person who opens it.

Save

Press Save in the bar at the top, or Ctrl+S. Problems show in the bar as a count — 2 to check — and on each step — No prompt set, No output key — this step's result is discarded, Unreachable from the entry point — so a half-built step is easy to find.

Give it an input

Open the arrow beside Run now in the bar at the top, type a question — What is the 30-day rule for shipping mail and internet orders? — and press Run. It runs the last saved version.

Watch it go

Each step lights up as it runs, and the run strip at the bottom left says where it is. Its Run Log follows the run, and the answer appears there when the run ends: a few sentences citing [K1] and [K3].

Read it back

Open Runs: the run is at the top. Its page shows each step's output and cost, the shared state the run left, and the exact prompt the model was sent. See Runs and what they cost.